Speech classification apparatus, speech classification method, and speech classification program

Abstract

A speech classification apparatus includes a speech classification probability calculation unit that calculates a probability (probability of classification into each cluster) that a latest one of the speech signals (speech data) belongs to each cluster based on a generative model which is a probability model, and a parameter updating unit that successively estimates parameters that define the generative model based on the probability of classification of the speech data into each cluster calculated by the speech classification probability calculation unit.

Description

[0001]The present invention is the National Phase of PCT / JP2008 / 054668 filed on Mar. 13, 2008, which claims the priority of Japanese Patent Application No. 2007-079677 (filed on Mar. 26, 2007), the entirety of which is incorporated herein by reference.TECHNICAL FIELD[0002]The present invention relates to a speech classification apparatus, a speech classification method, and a speech classification program that cluster speeches. More specifically, the invention relates to a speech classification apparatus, a speech classification method, and a speech classification program that cluster a plurality of speech signals issued by different speakers or in different environments into clusters based on similarity of the speakers or the environments.BACKGROUND ART[0003]As an example of a related art about the speech classification method, a description in Non-patent Document 1 is referred to, for example. FIG. 6 is a block diagram showing a configuration example of a speech classification app...

AI Technical Summary

Benefits of technology

The present invention provides a speech classification system that can operate in real-time and handle a large number of speeches with memory usage and processing time within reasonable limits. It also reduces classification errors and allows for verification and modification of classification errors when speeches are sequentially input. Additionally, it allows for more accurate classification of speeches based on detailed structures.

Problems solved by technology

A first problem is that the speech classification apparatuses and methods in the related arts are not suited to an application for which a real-time property is demanded.

That is, in such a method of calculating the distances between the two arbitrary speech data whenever one speech data is input, operation cannot be performed as an on-line process.

A second problem is that a memory use amount and a process time become enormous when a large quantity of speeches is handled.

A third problem is that, since greedy and deterministic classification is performed, a classification error may readily occur, and the error tends to adversely affect subsequent classification.

Thus, an overall group of speech data cannot be surveyed to find optimal classification.

A fourth problem is that, when speeches that are sequentially input on-line are classified, there is no simple means for reflecting a result of classification at a certain point of time on a preceding result of classification.

If a result of classification in the past is to be modified by the method described in Non-patent Document 2, almost the same method as that described in Non-patent Document 1 may be used, and the method is not suited to be carried out on-line.

A fifth problem is that accurate classification in view of the detailed structure of speech data cannot be performed.

Accordingly, accuracy of classification is naturally limited.

Images